An agent skills marketplace is a hosted directory where developers can publish, discover, install, and bundle reusable skills (capabilities, tools, prompts, or MCP servers) for AI coding agents. In the first week of June 2026, three vendors shipped on this thesis at the same time: Vercel made the skills.sh API generally available with 600,000 open-source skills, Cline released CLI 3.0.16 with plugin and skill bundling, and Anthropic announced Managed Agents with private MCP and 20+ vertical MCP connectors at its Code with Claude Tokyo event. This guide compares the five major agent skills marketplaces of 2026, explains how each one distributes skills, and shows where Totalum fits in for software agencies that need to ship production apps on top of them.
If you build for clients and want to turn agent skills into shippable production apps, book a 30-minute Totalum agency call for a live walkthrough.
Quick Answer
- The five marketplaces that matter in 2026: Anthropic Claude Skills, Vercel skills.sh, OpenAI Codex plugins, Cline plugins, and the MCP server ecosystem (Smithery, MCP Marketplace, LobeHub, SkillsMP).
- Vercel skills.sh went GA on June 5, 2026 with 600,000 OSS skills distributed via Vercel OIDC.
- Anthropic Claude Skills are markdown-defined capabilities bundled with code, callable from Claude Code, Claude Console, and Anthropic Managed Agents.
- The marketplaces differ on distribution, monetization, and lock-in. Totalum is the production-ship target across all five because it consumes skills via MCP and turns them into deployable Next.js apps with auth, payments, and storage.
- For software agencies running client work, the real question is not which marketplace to pick. It is which one to instrument first so your agents stop rewriting the same boilerplate every project.
What shipped in agent skills marketplaces this week (June 3 to 6, 2026)
A reactive snapshot you can quote.
- June 3, 2026. Cline released CLI 3.0.16 with a full plugin management system and skill bundling. Plugins can now ship as bundles with their dependencies and configuration. Slack socket support landed in the same release.
- June 4 to 5, 2026. Cursor 3.7 rolled out Organizations for Enterprise, Canvas Design Mode, and an SDK auto-review tool. Not a skills marketplace, but the SDK changes make Cursor a likely future entrant.
- June 5, 2026. Vercel made the skills.sh API generally available. The API exposes 600,000 open-source skills addressable via Vercel OIDC. Sandbox Drives (persistent storage) also entered private beta the same day.
- June 5 to 6, 2026. Anthropic at the Code with Claude Tokyo event announced Managed Agents that run inside customer-controlled sandboxes, connect to private MCP servers, and ship with 20+ legal-vertical MCP connectors plus 12 practice-area plugins. Claude Console is now available on AWS.
- June 5, 2026. Cline v3.88.0 made Fireworks Kimi K2.6 the default model and CLI 3.0.20 reinforced the plugin model.
Three vendors converged on the same theme in 72 hours. The marketplace era is real.
The five agent skills marketplaces compared
| Marketplace | First shipped | Hosted skills | Distribution | Monetization | Lock-in | Agency fit |
|---|---|---|---|---|---|---|
| Anthropic Claude Skills | October 2025 | Hundreds (curated) | Bundled with code via SKILL.md, callable from Claude Code, Claude Console, Managed Agents |
Free, included with Claude subscription | High (Anthropic-only runtime) | Strong if client commits to Claude |
| Vercel skills.sh | GA June 5, 2026 | 600,000 OSS skills | Vercel OIDC + skills.sh API, callable from any agent that signs in | Free OSS + paid Pro skills | Medium (uses Vercel OIDC but skills are portable) | Strong, especially for Next.js builds |
| OpenAI Codex plugins | May to June 2026 (default sharing) | Workspace-scoped + private | Workspace sharing, default plugin sharing in Enterprise / EDU | Free with ChatGPT Plus/Business | High (Codex runtime only) | Medium, workspace-bound |
| Cline plugins | June 3, 2026 (CLI 3.0.16) | Open ecosystem, self-hosted | Plugin bundles via Cline CLI, distributed through GitHub or registries | Free (OSS) | Low (Cline is OSS) | Strong if agency uses Cline locally |
| MCP server ecosystem (Smithery, LobeHub, MCP Marketplace, SkillsMP) | 2024 to 2026 | 425,000 (SkillsMP) + 169,000 (LobeHub) + thousands across registries | MCP protocol over stdio/SSE/HTTP, any MCP-capable client | Mostly free, some paid hosted | None (open protocol) | Strongest cross-vendor option |
Sources: Vercel changelog (skills.sh API GA, June 5, 2026), Cline GitHub releases (v3.87.0 to v3.88.0, June 3 to 5, 2026), Anthropic news (Code with Claude Tokyo, June 5 to 6, 2026), Cursor changelog (3.7, June 3 to 5, 2026), SkillsMP / LobeHub registry counts (2026 estimates).
How each marketplace works
Anthropic Claude Skills
A Claude Skill is a folder containing a SKILL.md file plus the code or assets it needs. The SKILL.md is a markdown spec the model reads to decide when to invoke the skill and what arguments to pass. Claude Code, Claude Console, and Managed Agents all consume the same skill format, which means a skill you write for a desktop coding session works inside a server-side agent without changes.
Anthropic's marketplace is curated rather than open submission. Skills ship through the Claude Skills directory and a growing list of partner repositories. The June 5 to 6, 2026 Tokyo announcements added 20+ MCP connectors for legal verticals and 12 practice-area plugins (e.g., contract review, IP filings) as managed bundles.
Best for teams already on the Claude stack. The cost is portability: a Claude Skill is hard to run inside a Codex workspace or a Cline session without rewriting it.
Vercel skills.sh
skills.sh is Vercel's bet that skills should be addressable like packages. The skills.sh API GA on June 5, 2026 exposed 600,000 open-source skills, each callable by name with Vercel OIDC handling auth. A skill on skills.sh looks like an HTTP endpoint plus a JSON manifest that any agent runtime can consume.
The distribution model is open-source first. Anyone can publish a skill; Vercel scores and ranks them; paid Pro skills run on Vercel's managed infrastructure with billing handled through Vercel's account. Sandbox Drives (private beta the same day) add persistent storage for skills that need state across runs.
Best for teams already deployed on Vercel and for skills that benefit from being callable across multiple agents. The lock-in is moderate: the auth layer is Vercel's, but the skills themselves are portable.
OpenAI Codex plugins
Codex plugins are workspace-scoped capabilities that extend a Codex coding session. The June 4 to 5, 2026 releases (Codex 26.602 and 0.137.0) added enterprise flows, plugin workflow tooling, and default plugin sharing in eligible Enterprise / EDU workspaces.
Plugins are distributed inside an OpenAI workspace. A plugin you publish to your team's workspace can be installed by every member; default sharing in Enterprise/EDU workspaces means new members inherit the workspace's plugin library on day one.
Best for teams standardized on ChatGPT Enterprise. The cost is reach: a Codex plugin is hard to expose outside the workspace, and there is no public marketplace as of June 2026.
Cline plugins (CLI 3.0.16)
Cline's June 3, 2026 release made plugins first-class. A Cline plugin is a bundle with its dependencies, configuration, and entry points; the CLI manages installation, updates, and runtime. The same release added Slack socket support, which means an agent can publish results to a channel without bespoke wiring.
Cline is open source, so the plugin ecosystem is decentralized. Plugins are typically distributed through GitHub repositories or public registries; teams self-host the runtime. There is no central monetization layer.
Best for engineering teams that want full control of the runtime and are willing to maintain their own plugin registry. The lock-in is the lowest of the five.
MCP server ecosystem (Smithery, MCP Marketplace, LobeHub, SkillsMP)
MCP is an open protocol. Any MCP server can be consumed by any MCP-capable client (Claude Desktop, Cursor, Claude Code, Cline, custom agents). The result is a sprawling cross-vendor marketplace: SkillsMP lists 425,000+ skills, LobeHub lists 169,000+, and Smithery plus the MCP Marketplace index thousands of servers across hosting models.
Distribution is whatever the server author chooses (stdio, SSE, HTTP). Monetization is mostly free, with hosted offerings (e.g., Smithery's managed servers) charging on usage. Lock-in is effectively zero because the protocol is open.
Best as the default for teams that want to avoid betting on a single vendor. The cost is curation: with hundreds of thousands of servers across registries, choosing the right one is its own job. Our reference guides on the best MCP servers in 2026 and Claude Code MCP servers walk through what to install first.
How to evaluate an agent skills marketplace before you commit
Before you wire your agency or your SaaS to a single marketplace, run every candidate through this six-point checklist. Most teams that get stuck on a marketplace 18 months in skipped at least one of these checks at adoption time.
- Portability. Can a skill written for this marketplace run inside another agent? Anthropic Claude Skills and OpenAI Codex plugins fail this check by design. Cline plugins and MCP servers pass it. skills.sh is somewhere in the middle because its manifest format is HTTP-callable but not native to other runtimes.
- Distribution friction. How does a new skill get discovered? Curated marketplaces (Anthropic, OpenAI) gate discovery on the vendor's editorial process. Open ones (Cline, MCP) push discovery onto the user, which is why curated lists like the one in our Claude Code vs Codex 2026 comparison have outsized impact.
- Auth surface. Vercel skills.sh ships with OIDC and turns auth into a single managed concern. MCP servers leave auth to the server author. For client work, fewer auth layers is fewer ways the agent breaks at 4 PM on a Friday.
- Versioning. Anthropic Skills version through git. Cline bundles version through the CLI. skills.sh versions through the skill manifest. Pick a marketplace whose versioning matches how your team already ships.
- Failure mode. When a skill fails inside an agent run, what does the agent see? Marketplaces that return structured errors (Vercel skills.sh, MCP) are easier to debug than marketplaces that return free-form text.
- Monetization clarity. If you plan to publish skills you charge for, only skills.sh and the hosted MCP layer (Smithery, etc.) have a clear billing path today. Anthropic and OpenAI marketplaces will likely add billing in 2026 to 2027, but as of June 2026 there is no native payments rail.
If you are building an AI agent platform on top of Totalum, the same checklist applies in reverse: every skill you expose to your customers is a marketplace listing in their workflow, and they will judge it on the same six axes.
Why Totalum fits the agent skills marketplace era
The five marketplaces above are about distributing capabilities. Totalum is about turning those capabilities into shipped products. That makes Totalum the production-ship target across all five.
Totalum is an AI app builder for humans and for agents. It produces real, deployable Next.js applications with built-in auth, payments, database, file storage, and AI integrations. The same builder is callable via API and MCP, which means any agent in any of the five marketplaces above can talk to Totalum and ship a working app on a custom domain.
For a software agency, the practical workflow is:
- The client asks for an internal tool, a SaaS feature, or a portal.
- The agency's coding agent (Claude Code, Cursor, Codex, Cline) loads the skills it needs from the marketplace of choice (e.g., a Claude Skill for PDF parsing plus an MCP server for the client's CRM).
- The agent calls Totalum's MCP server to scaffold the app, write the schema, hook up auth, configure payments, and deploy to a custom domain.
- The agency ships the production app the same day, billed as a fixed engagement.
The marketplaces accelerate the agent's reasoning. Totalum accelerates the agent's shipping. The two are complementary, and an agency that wires both layers can deliver client work at a multiple of the rate a traditional shop can.
For SaaS founders who want to embed a builder inside their own product, the same MCP surface is what powers a "build your own Lovable" feature. See our guide on how to embed an AI app builder into your SaaS via API for the integration pattern.
FAQ
Which agent skills marketplace should an agency pick first?
If your team is already standardized on a single coding agent, pick its native marketplace first (Claude Skills for Claude shops, Codex plugins for ChatGPT Enterprise shops, Cline plugins for OSS-first teams). Then layer MCP servers on top because they work across every agent. Avoid betting only on a vendor-locked marketplace if your clients use different tools.
How is Vercel skills.sh different from the MCP server ecosystem?
skills.sh is a single hosted directory addressable via Vercel OIDC. The MCP server ecosystem is an open protocol with multiple registries (Smithery, LobeHub, SkillsMP, MCP Marketplace). A skill on skills.sh is consumed by HTTP plus a JSON manifest; an MCP server speaks the MCP protocol (stdio, SSE, or HTTP). The two can coexist: a skills.sh skill can wrap an MCP call, and an MCP server can expose a skills.sh-hosted capability.
Are Claude Skills the same as Claude Code subagents and hooks?
No. Claude Skills are reusable capabilities defined by a SKILL.md file plus code. Subagents are spawned worker contexts inside a Claude Code session. Hooks are event handlers that fire on lifecycle events (pre-tool, post-tool, etc.). Skills are what you call; subagents are who calls them; hooks are when they fire. Our Claude Skills + Totalum guide covers the distinctions in depth.
Can I monetize a skill I publish to one of these marketplaces?
Vercel skills.sh has a paid Pro tier for hosted skills. The MCP ecosystem has hosted-server billing through Smithery and similar registries. Anthropic Claude Skills and OpenAI Codex plugins are currently free distribution channels with no native billing layer; monetization happens outside the marketplace (subscription, services, enterprise contract).
How do I keep agent skills secure when they run with production credentials?
Treat every skill as an untrusted dependency. Audit the SKILL.md or manifest before installing, run skills in sandboxed environments (Anthropic Managed Agents, Vercel Sandbox Drives, Cline's local runtime), and rotate any credentials the skill touches. For production apps built on Totalum, secrets are stored in the project's secrets vault and never exposed to the agent unless the skill explicitly requests them.
What is the difference between using Anthropic Skills and using Totalum directly?
Anthropic Skills are capabilities your Claude session can call during reasoning. Totalum is the production-ship target where the resulting app actually runs. A common pattern: Claude Code loads three Skills from Anthropic's marketplace, runs them to gather context and process data, then calls Totalum's MCP to deploy the final app. They are not competitors; they sit at different layers of the stack.
Ready to ship on top of the agent skills marketplace?
If your agency wants to instrument these marketplaces and turn agent output into shipped client apps in days, not months, book a 30-minute Totalum discovery call. We will walk through your current stack, show how Totalum's MCP slots into your agent of choice, and quote a pilot project.
For SaaS founders embedding an AI builder into their own product, the same call covers the embedding pattern via Totalum's API/MCP.
If you would rather just try it, start free at totalum.app and connect your agent to Totalum's MCP in under 10 minutes. Totalum is the agent skills marketplace era's production-ready destination, and we are pricing the agency tier accordingly while the window is open.